Objective:

To characterize the frequency, timing, nature, and FDA documentation of updates to AI/ML-based medical devices cleared through the 510(k) pathway.

Approach:

• Study design: Cross-sectional analysis of a previously identified cohort of AI/ML-based devices cleared from January 9, 2019, through December 29, 2021; FDA database entries were reviewed through March 31, 2025.
• Update assessment: Reviewers identified modification requests involving algorithm updates, classified changes by whether they altered device use or improved algorithm performance, and assessed reporting of performance evaluations and benefits.

Key Findings:

• The analysis included 163 devices and 352 FDA database entries; 89.0% of devices analyzed medical imaging data.
• Fifty-six devices (34.4%) had at least one identified AI/ML algorithm update. The median time to the first update was 1.8 years (IQR, 0.9–2.5).
• Among updated devices, 42 (75.0%) had at least one update that altered use. Changes included new features and changes to algorithm inputs.
• Performance-only updates were identified for 9 devices (16.1%); the update method was unclear for 5 devices (8.9%).
• FDA documentation of update evaluations was incomplete: participant information was absent from 38.7% of update entries, sample size from 35.8%, and performance metrics and results from 13.2%. Information on organizational or clinical benefits was absent from 86 of 106 entries (81.1%); 48 of 56 devices (85.7%) had no such information in any update-related entry.

Interpretation:

The authors reported that updates were common among the studied devices and that FDA documentation varied considerably in its reporting of modifications and associated performance. They noted that incomplete reporting may limit physicians’ and patients’ ability to evaluate and select these devices.

Limitations:

• The study used a previously identified device cohort and included a relatively small sample.
• Classification may have been inaccurate because it relied exclusively on FDA documents.
• The analysis could not determine whether devices without identified updates had been modified without an associated FDA modification request.
• Not all updates were captured; some changes may be made under FDA Predetermined Change Control Plan guidance without a new application.

Conclusion:

The authors stated that enhancing and standardizing reporting requirements, alongside establishing and enforcing postdeployment monitoring, is important for ongoing assessment of AI/ML-based medical device safety and effectiveness.